This project is a key component of our PtrSplit and Program-mandering works. It aims at building a modular inter-procedural program dependence graph (PDG) for practical use. Our program dependence graph is field senstive, context-insensitive and flow-insensitive. For more details, welcome to read our CCS'17 paper about PtrSplit: [http://www.cse.psu.edu/~gxt29/papers/ptrsplit.pdf] If you find this tool useful, please cite the PtrSplit and Program Mandering papers in your publication. Here's the bibtex entries:

@inproceedings{LiuTJ17Ptrsplit, author = {Shen Liu and Gang Tan and Trent Jaeger}, title = {{PtrSplit}: Supporting General Pointers in Automatic Program Partitioning}, booktitle = {24th ACM Conference on Computer and Communications Security ({CCS})}, pages = {2359--2371}, year = {2017} }

@inproceedings{liu2019program, title={Program-mandering: Quantitative privilege separation}, author={Liu, Shen and Zeng, Dongrui and Huang, Yongzhe and Capobianco, Frank and McCamant, Stephen and Jaeger, Trent and Tan, Gang}, booktitle={Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security}, pages={1023--1040}, year={2019} }

We have upgraded the implementation to LLVM 12.0.0. Currently, we only support building PDGs for C programs.

A PDG example looks like this (the blue part corresponds to the parameter tree):

mkdir build
cd build
cmake ..
make
opt -load libpdg.so -dot-pdg < test.bc

-pdg: generate the program dependence graph (inter-procedural)

-cdg: generate the control dependence graph (intra-procedural)

-ddg: generate the data dependence graph (intra-procedural)

-dot-*: for visualization. (dot)

For those large software, generating a visualizable PDG is not easy. Graphviz often fails to generate the .dot file for a program with more than 1000 lines of C code. Fortunately, we rarely need such a large .dot file but only do kinds of analyses on the PDG, which is always in memory.

For simple C programs(e.g., test.c), do

clang -emit-llvm -S -g test.c -o test.bc

Now you have a binary format LLVM bitcode file which can be directly used as the input for PDG generation.

For those large C software (e.g., wget), you can refer to this great article for help:

Compiling Autotooled projects to LLVM Bitcode

(We successfully compiled SPECCPU 2006 INT/thttpd/wget/telnet/openssh/curl/nginx/sqlite, thanks to the author!)

We can use the current PDG as a required pass through following steps:

1. download PDG repo: git clone https://github.com/ARISTODE/program-dependence-graph.git
2. cd program-dependence-graph
3. make

Using cmake, add

include_directories(program_dependence_graph/include)
add_subdirectory(program_dependence_graph)

Then, add

AU.addRequired<ProgramDependencyGraph>();

in your pass's getAnalysisUsage method (legacy pass manager).

Query the reachability of two nodes:

ProgramGraph *g = getAnalysis<ProgramDependencyGraph>()->getPDG();

Value* src;
Value* dst;

pdg::Node* src_node = g->getNode(*src);
pdg::Node* dst_node = g->getNode(*dst);

if (g->canReach(src_node, dst_node)) 
{
  // do something...
}

Traverse the PDG with path constrains This method is useful to traverse the graph through certain edge types. In the example, we put the edge types we want to exclude in the set exclude_edges. Then, pass that as an argument to the canReach function.

ProgramGraph *g = getAnalysis<ProgramDependencyGraph>()->getPDG();

Value* src;
Value* dst;

pdg::Node* src_node = g->getNode(*src);
pdg::Node* dst_node = g->getNode(*dst);

std::set<pdg::EdgeType> exclude_edges;

if (g->canReach(src_node, dst_node, exclude_edges)) 
{
  // do something...
}